Supports lookup devices by fullname either in the canonical form or the

legacy form. This makes DeviceSet behaves the same as DeviceMgr's
FindDevice method.

PiperOrigin-RevId: 163300346

tensorflow/core/common_runtime/device_mgr.cc

@@ -30,22 +30,9 @@ DeviceMgr::DeviceMgr(const std::vector<Device*>& devices)
     devices_.push_back(d);
 
     // Register under the (1) full name, (2) canonical name, and (3) local name.
-    string full_name = d->name();
-    device_map_[CopyToBackingStore(full_name)] = d;
-
-    // TODO(b/62909072): Upgrade device_map_ to a better data structure.
-    DeviceNameUtils::ParsedName parsed_name = d->parsed_name();
-    if (parsed_name.has_job && parsed_name.has_replica &&
-        parsed_name.has_task && parsed_name.has_type && parsed_name.has_id) {
-      string canonical_name = DeviceNameUtils::FullName(
-          parsed_name.job, parsed_name.replica, parsed_name.task,
-          parsed_name.type, parsed_name.id);
-      device_map_[CopyToBackingStore(canonical_name)] = d;
-
-      string legacy_name = DeviceNameUtils::LegacyName(
-          parsed_name.job, parsed_name.replica, parsed_name.task,
-          parsed_name.type, parsed_name.id);
-      device_map_[CopyToBackingStore(legacy_name)] = d;
+    for (const string& name :
+         DeviceNameUtils::GetNamesForDeviceMappings(d->parsed_name())) {
+      device_map_[CopyToBackingStore(name)] = d;
     }
     string lname = DeviceNameUtils::LocalName(d->name());
     device_map_[CopyToBackingStore(lname)] = d;

tensorflow/core/common_runtime/device_set.cc

@@ -32,7 +32,10 @@ DeviceSet::~DeviceSet() {}
 
 void DeviceSet::AddDevice(Device* device) {
   devices_.push_back(device);
-  device_by_name_.insert({device->name(), device});
+  for (const string& name :
+       DeviceNameUtils::GetNamesForDeviceMappings(device->parsed_name())) {
+    device_by_name_.insert({name, device});
+  }
 }
 
 void DeviceSet::FindMatchingDevices(const DeviceNameUtils::ParsedName& spec,

tensorflow/core/util/device_name_utils.cc

@@ -387,4 +387,16 @@ bool DeviceNameUtils::SplitDeviceName(StringPiece name, string* task,
   return false;
 }
 
+std::vector<string> DeviceNameUtils::GetNamesForDeviceMappings(
+    const ParsedName& pn) {
+  if (pn.has_job && pn.has_replica && pn.has_task && pn.has_type && pn.has_id) {
+    return {
+        DeviceNameUtils::FullName(pn.job, pn.replica, pn.task, pn.type, pn.id),
+        DeviceNameUtils::LegacyName(pn.job, pn.replica, pn.task, pn.type,
+                                    pn.id)};
+  } else {
+    return {};
+  }
+}
+
 }  // namespace tensorflow

tensorflow/core/util/device_name_utils.h

@@ -152,6 +152,11 @@ class DeviceNameUtils {
   static bool SplitDeviceName(StringPiece name, string* task, string* device);
 
   static string ParsedNameToString(const ParsedName& pn);
+
+  // Returns canonical and legacy full names for the given parsed
+  // device name 'pn'. The returned string names are often useful to
+  // lookup devices from a mapping.
+  static std::vector<string> GetNamesForDeviceMappings(const ParsedName& pn);
 };
 
 }  // namespace tensorflow

tensorflow/core/util/device_name_utils_test.cc

@@ -443,6 +443,16 @@ TEST(DeviceNameUtilsTest, MergeDevNamesAllowSoftPlacement) {
   MergeDevNamesHelperAllowSoftPlacement("/gpu:1", "/gpu:2", "/gpu:*");
 }
 
+TEST(DeviceNameUtilsTest, GetNamesForDeviceMappings) {
+  DeviceNameUtils::ParsedName p = Name("/job:foo/replica:10/task:0/gpu:1");
+  EXPECT_EQ(str_util::Join(DeviceNameUtils::GetNamesForDeviceMappings(p), ","),
+            "/job:foo/replica:10/task:0/device:GPU:1,"
+            "/job:foo/replica:10/task:0/gpu:1");
+  p.has_task = false;
+  EXPECT_EQ(str_util::Join(DeviceNameUtils::GetNamesForDeviceMappings(p), ","),
+            "");
+}
+
 static void BM_ParseFullName(int iters) {
   DeviceNameUtils::ParsedName p;
   while (iters--) {